mm/userfaultfd.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  mm/userfaultfd.c
   4  *
   5  *  Copyright (C) 2015  Red Hat, Inc.
   6  */
   7
   8 #include <linux/mm.h>
   9 #include <linux/sched/signal.h>
  10 #include <linux/pagemap.h>
  11 #include <linux/rmap.h>
  12 #include <linux/swap.h>
  13 #include <linux/swapops.h>
  14 #include <linux/userfaultfd_k.h>
  15 #include <linux/mmu_notifier.h>
  16 #include <linux/hugetlb.h>
  17 #include <linux/shmem_fs.h>
  18 #include <asm/tlbflush.h>
  19 #include <asm/tlb.h>
  20 #include "internal.h"
  21
  22 static __always_inline
  23 struct vm_area_struct *find_dst_vma(struct mm_struct *dst_mm,
  24                                     unsigned long dst_start,
  25                                     unsigned long len)
  26 {
  27         /*
  28          * Make sure that the dst range is both valid and fully within a
  29          * single existing vma.
  30          */
  31         struct vm_area_struct *dst_vma;
  32
  33         dst_vma = find_vma(dst_mm, dst_start);
  34         if (!dst_vma)
  35                 return NULL;
  36
  37         if (dst_start < dst_vma->vm_start ||
  38             dst_start + len > dst_vma->vm_end)
  39                 return NULL;
  40
  41         /*
  42          * Check the vma is registered in uffd, this is required to
  43          * enforce the VM_MAYWRITE check done at uffd registration
  44          * time.
  45          */
  46         if (!dst_vma->vm_userfaultfd_ctx.ctx)
  47                 return NULL;
  48
  49         return dst_vma;
  50 }
  51
  52 /*
  53  * Install PTEs, to map dst_addr (within dst_vma) to page.
  54  *
  55  * This function handles both MCOPY_ATOMIC_NORMAL and _CONTINUE for both shmem
  56  * and anon, and for both shared and private VMAs.
  57  */
  58 int mfill_atomic_install_pte(struct mm_struct *dst_mm, pmd_t *dst_pmd,
  59                              struct vm_area_struct *dst_vma,
  60                              unsigned long dst_addr, struct page *page,
  61                              bool newly_allocated, bool wp_copy)
  62 {
  63         int ret;
  64         pte_t _dst_pte, *dst_pte;
  65         bool writable = dst_vma->vm_flags & VM_WRITE;
  66         bool vm_shared = dst_vma->vm_flags & VM_SHARED;
  67         bool page_in_cache = page_mapping(page);
  68         spinlock_t *ptl;
  69         struct inode *inode;
  70         pgoff_t offset, max_off;
  71
  72         _dst_pte = mk_pte(page, dst_vma->vm_page_prot);
  73         _dst_pte = pte_mkdirty(_dst_pte);
  74         if (page_in_cache && !vm_shared)
  75                 writable = false;
  76
  77         /*
  78          * Always mark a PTE as write-protected when needed, regardless of
  79          * VM_WRITE, which the user might change.
  80          */
  81         if (wp_copy) {
  82                 _dst_pte = pte_mkuffd_wp(_dst_pte);
  83                 writable = false;
  84         }
  85
  86         if (writable)
  87                 _dst_pte = pte_mkwrite(_dst_pte);
  88         else
  89                 /*
  90                  * We need this to make sure write bit removed; as mk_pte()
  91                  * could return a pte with write bit set.
  92                  */
  93                 _dst_pte = pte_wrprotect(_dst_pte);
  94
  95         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
  96
  97         if (vma_is_shmem(dst_vma)) {
  98                 /* serialize against truncate with the page table lock */
  99                 inode = dst_vma->vm_file->f_inode;
 100                 offset = linear_page_index(dst_vma, dst_addr);
 101                 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
 102                 ret = -EFAULT;
 103                 if (unlikely(offset >= max_off))
 104                         goto out_unlock;
 105         }
 106
 107         ret = -EEXIST;
 108         /*
 109          * We allow to overwrite a pte marker: consider when both MISSING|WP
 110          * registered, we firstly wr-protect a none pte which has no page cache
 111          * page backing it, then access the page.
 112          */
 113         if (!pte_none_mostly(*dst_pte))
 114                 goto out_unlock;
 115
 116         if (page_in_cache) {
 117                 /* Usually, cache pages are already added to LRU */
 118                 if (newly_allocated)
 119                         lru_cache_add(page);
 120                 page_add_file_rmap(page, dst_vma, false);
 121         } else {
 122                 page_add_new_anon_rmap(page, dst_vma, dst_addr);
 123                 lru_cache_add_inactive_or_unevictable(page, dst_vma);
 124         }
 125
 126         /*
 127          * Must happen after rmap, as mm_counter() checks mapping (via
 128          * PageAnon()), which is set by __page_set_anon_rmap().
 129          */
 130         inc_mm_counter(dst_mm, mm_counter(page));
 131
 132         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
 133
 134         /* No need to invalidate - it was non-present before */
 135         update_mmu_cache(dst_vma, dst_addr, dst_pte);
 136         ret = 0;
 137 out_unlock:
 138         pte_unmap_unlock(dst_pte, ptl);
 139         return ret;
 140 }
 141
 142 static int mcopy_atomic_pte(struct mm_struct *dst_mm,
 143                             pmd_t *dst_pmd,
 144                             struct vm_area_struct *dst_vma,
 145                             unsigned long dst_addr,
 146                             unsigned long src_addr,
 147                             struct page **pagep,
 148                             bool wp_copy)
 149 {
 150         void *page_kaddr;
 151         int ret;
 152         struct page *page;
 153
 154         if (!*pagep) {
 155                 ret = -ENOMEM;
 156                 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, dst_vma, dst_addr);
 157                 if (!page)
 158                         goto out;
 159
 160                 page_kaddr = kmap_local_page(page);
 161                 /*
 162                  * The read mmap_lock is held here.  Despite the
 163                  * mmap_lock being read recursive a deadlock is still
 164                  * possible if a writer has taken a lock.  For example:
 165                  *
 166                  * process A thread 1 takes read lock on own mmap_lock
 167                  * process A thread 2 calls mmap, blocks taking write lock
 168                  * process B thread 1 takes page fault, read lock on own mmap lock
 169                  * process B thread 2 calls mmap, blocks taking write lock
 170                  * process A thread 1 blocks taking read lock on process B
 171                  * process B thread 1 blocks taking read lock on process A
 172                  *
 173                  * Disable page faults to prevent potential deadlock
 174                  * and retry the copy outside the mmap_lock.
 175                  */
 176                 pagefault_disable();
 177                 ret = copy_from_user(page_kaddr,
 178                                      (const void __user *) src_addr,
 179                                      PAGE_SIZE);
 180                 pagefault_enable();
 181                 kunmap_local(page_kaddr);
 182
 183                 /* fallback to copy_from_user outside mmap_lock */
 184                 if (unlikely(ret)) {
 185                         ret = -ENOENT;
 186                         *pagep = page;
 187                         /* don't free the page */
 188                         goto out;
 189                 }
 190
 191                 flush_dcache_page(page);
 192         } else {
 193                 page = *pagep;
 194                 *pagep = NULL;
 195         }
 196
 197         /*
 198          * The memory barrier inside __SetPageUptodate makes sure that
 199          * preceding stores to the page contents become visible before
 200          * the set_pte_at() write.
 201          */
 202         __SetPageUptodate(page);
 203
 204         ret = -ENOMEM;
 205         if (mem_cgroup_charge(page_folio(page), dst_mm, GFP_KERNEL))
 206                 goto out_release;
 207
 208         ret = mfill_atomic_install_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 209                                        page, true, wp_copy);
 210         if (ret)
 211                 goto out_release;
 212 out:
 213         return ret;
 214 out_release:
 215         put_page(page);
 216         goto out;
 217 }
 218
 219 static int mfill_zeropage_pte(struct mm_struct *dst_mm,
 220                               pmd_t *dst_pmd,
 221                               struct vm_area_struct *dst_vma,
 222                               unsigned long dst_addr)
 223 {
 224         pte_t _dst_pte, *dst_pte;
 225         spinlock_t *ptl;
 226         int ret;
 227         pgoff_t offset, max_off;
 228         struct inode *inode;
 229
 230         _dst_pte = pte_mkspecial(pfn_pte(my_zero_pfn(dst_addr),
 231                                          dst_vma->vm_page_prot));
 232         dst_pte = pte_offset_map_lock(dst_mm, dst_pmd, dst_addr, &ptl);
 233         if (dst_vma->vm_file) {
 234                 /* the shmem MAP_PRIVATE case requires checking the i_size */
 235                 inode = dst_vma->vm_file->f_inode;
 236                 offset = linear_page_index(dst_vma, dst_addr);
 237                 max_off = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
 238                 ret = -EFAULT;
 239                 if (unlikely(offset >= max_off))
 240                         goto out_unlock;
 241         }
 242         ret = -EEXIST;
 243         if (!pte_none(*dst_pte))
 244                 goto out_unlock;
 245         set_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
 246         /* No need to invalidate - it was non-present before */
 247         update_mmu_cache(dst_vma, dst_addr, dst_pte);
 248         ret = 0;
 249 out_unlock:
 250         pte_unmap_unlock(dst_pte, ptl);
 251         return ret;
 252 }
 253
 254 /* Handles UFFDIO_CONTINUE for all shmem VMAs (shared or private). */
 255 static int mcontinue_atomic_pte(struct mm_struct *dst_mm,
 256                                 pmd_t *dst_pmd,
 257                                 struct vm_area_struct *dst_vma,
 258                                 unsigned long dst_addr,
 259                                 bool wp_copy)
 260 {
 261         struct inode *inode = file_inode(dst_vma->vm_file);
 262         pgoff_t pgoff = linear_page_index(dst_vma, dst_addr);
 263         struct folio *folio;
 264         struct page *page;
 265         int ret;
 266
 267         ret = shmem_get_folio(inode, pgoff, &folio, SGP_NOALLOC);
 268         /* Our caller expects us to return -EFAULT if we failed to find folio */
 269         if (ret == -ENOENT)
 270                 ret = -EFAULT;
 271         if (ret)
 272                 goto out;
 273         if (!folio) {
 274                 ret = -EFAULT;
 275                 goto out;
 276         }
 277
 278         page = folio_file_page(folio, pgoff);
 279         if (PageHWPoison(page)) {
 280                 ret = -EIO;
 281                 goto out_release;
 282         }
 283
 284         ret = mfill_atomic_install_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 285                                        page, false, wp_copy);
 286         if (ret)
 287                 goto out_release;
 288
 289         folio_unlock(folio);
 290         ret = 0;
 291 out:
 292         return ret;
 293 out_release:
 294         folio_unlock(folio);
 295         folio_put(folio);
 296         goto out;
 297 }
 298
 299 static pmd_t *mm_alloc_pmd(struct mm_struct *mm, unsigned long address)
 300 {
 301         pgd_t *pgd;
 302         p4d_t *p4d;
 303         pud_t *pud;
 304
 305         pgd = pgd_offset(mm, address);
 306         p4d = p4d_alloc(mm, pgd, address);
 307         if (!p4d)
 308                 return NULL;
 309         pud = pud_alloc(mm, p4d, address);
 310         if (!pud)
 311                 return NULL;
 312         /*
 313          * Note that we didn't run this because the pmd was
 314          * missing, the *pmd may be already established and in
 315          * turn it may also be a trans_huge_pmd.
 316          */
 317         return pmd_alloc(mm, pud, address);
 318 }
 319
 320 #ifdef CONFIG_HUGETLB_PAGE
 321 /*
 322  * __mcopy_atomic processing for HUGETLB vmas.  Note that this routine is
 323  * called with mmap_lock held, it will release mmap_lock before returning.
 324  */
 325 static __always_inline ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 326                                               struct vm_area_struct *dst_vma,
 327                                               unsigned long dst_start,
 328                                               unsigned long src_start,
 329                                               unsigned long len,
 330                                               enum mcopy_atomic_mode mode,
 331                                               bool wp_copy)
 332 {
 333         int vm_shared = dst_vma->vm_flags & VM_SHARED;
 334         ssize_t err;
 335         pte_t *dst_pte;
 336         unsigned long src_addr, dst_addr;
 337         long copied;
 338         struct page *page;
 339         unsigned long vma_hpagesize;
 340         pgoff_t idx;
 341         u32 hash;
 342         struct address_space *mapping;
 343
 344         /*
 345          * There is no default zero huge page for all huge page sizes as
 346          * supported by hugetlb.  A PMD_SIZE huge pages may exist as used
 347          * by THP.  Since we can not reliably insert a zero page, this
 348          * feature is not supported.
 349          */
 350         if (mode == MCOPY_ATOMIC_ZEROPAGE) {
 351                 mmap_read_unlock(dst_mm);
 352                 return -EINVAL;
 353         }
 354
 355         src_addr = src_start;
 356         dst_addr = dst_start;
 357         copied = 0;
 358         page = NULL;
 359         vma_hpagesize = vma_kernel_pagesize(dst_vma);
 360
 361         /*
 362          * Validate alignment based on huge page size
 363          */
 364         err = -EINVAL;
 365         if (dst_start & (vma_hpagesize - 1) || len & (vma_hpagesize - 1))
 366                 goto out_unlock;
 367
 368 retry:
 369         /*
 370          * On routine entry dst_vma is set.  If we had to drop mmap_lock and
 371          * retry, dst_vma will be set to NULL and we must lookup again.
 372          */
 373         if (!dst_vma) {
 374                 err = -ENOENT;
 375                 dst_vma = find_dst_vma(dst_mm, dst_start, len);
 376                 if (!dst_vma || !is_vm_hugetlb_page(dst_vma))
 377                         goto out_unlock;
 378
 379                 err = -EINVAL;
 380                 if (vma_hpagesize != vma_kernel_pagesize(dst_vma))
 381                         goto out_unlock;
 382
 383                 vm_shared = dst_vma->vm_flags & VM_SHARED;
 384         }
 385
 386         /*
 387          * If not shared, ensure the dst_vma has a anon_vma.
 388          */
 389         err = -ENOMEM;
 390         if (!vm_shared) {
 391                 if (unlikely(anon_vma_prepare(dst_vma)))
 392                         goto out_unlock;
 393         }
 394
 395         while (src_addr < src_start + len) {
 396                 BUG_ON(dst_addr >= dst_start + len);
 397
 398                 /*
 399                  * Serialize via vma_lock and hugetlb_fault_mutex.
 400                  * vma_lock ensures the dst_pte remains valid even
 401                  * in the case of shared pmds.  fault mutex prevents
 402                  * races with other faulting threads.
 403                  */
 404                 idx = linear_page_index(dst_vma, dst_addr);
 405                 mapping = dst_vma->vm_file->f_mapping;
 406                 hash = hugetlb_fault_mutex_hash(mapping, idx);
 407                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 408                 hugetlb_vma_lock_read(dst_vma);
 409
 410                 err = -ENOMEM;
 411                 dst_pte = huge_pte_alloc(dst_mm, dst_vma, dst_addr, vma_hpagesize);
 412                 if (!dst_pte) {
 413                         hugetlb_vma_unlock_read(dst_vma);
 414                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 415                         goto out_unlock;
 416                 }
 417
 418                 if (mode != MCOPY_ATOMIC_CONTINUE &&
 419                     !huge_pte_none_mostly(huge_ptep_get(dst_pte))) {
 420                         err = -EEXIST;
 421                         hugetlb_vma_unlock_read(dst_vma);
 422                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 423                         goto out_unlock;
 424                 }
 425
 426                 err = hugetlb_mcopy_atomic_pte(dst_mm, dst_pte, dst_vma,
 427                                                dst_addr, src_addr, mode, &page,
 428                                                wp_copy);
 429
 430                 hugetlb_vma_unlock_read(dst_vma);
 431                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 432
 433                 cond_resched();
 434
 435                 if (unlikely(err == -ENOENT)) {
 436                         mmap_read_unlock(dst_mm);
 437                         BUG_ON(!page);
 438
 439                         err = copy_huge_page_from_user(page,
 440                                                 (const void __user *)src_addr,
 441                                                 vma_hpagesize / PAGE_SIZE,
 442                                                 true);
 443                         if (unlikely(err)) {
 444                                 err = -EFAULT;
 445                                 goto out;
 446                         }
 447                         mmap_read_lock(dst_mm);
 448
 449                         dst_vma = NULL;
 450                         goto retry;
 451                 } else
 452                         BUG_ON(page);
 453
 454                 if (!err) {
 455                         dst_addr += vma_hpagesize;
 456                         src_addr += vma_hpagesize;
 457                         copied += vma_hpagesize;
 458
 459                         if (fatal_signal_pending(current))
 460                                 err = -EINTR;
 461                 }
 462                 if (err)
 463                         break;
 464         }
 465
 466 out_unlock:
 467         mmap_read_unlock(dst_mm);
 468 out:
 469         if (page)
 470                 put_page(page);
 471         BUG_ON(copied < 0);
 472         BUG_ON(err > 0);
 473         BUG_ON(!copied && !err);
 474         return copied ? copied : err;
 475 }
 476 #else /* !CONFIG_HUGETLB_PAGE */
 477 /* fail at build time if gcc attempts to use this */
 478 extern ssize_t __mcopy_atomic_hugetlb(struct mm_struct *dst_mm,
 479                                       struct vm_area_struct *dst_vma,
 480                                       unsigned long dst_start,
 481                                       unsigned long src_start,
 482                                       unsigned long len,
 483                                       enum mcopy_atomic_mode mode,
 484                                       bool wp_copy);
 485 #endif /* CONFIG_HUGETLB_PAGE */
 486
 487 static __always_inline ssize_t mfill_atomic_pte(struct mm_struct *dst_mm,
 488                                                 pmd_t *dst_pmd,
 489                                                 struct vm_area_struct *dst_vma,
 490                                                 unsigned long dst_addr,
 491                                                 unsigned long src_addr,
 492                                                 struct page **page,
 493                                                 enum mcopy_atomic_mode mode,
 494                                                 bool wp_copy)
 495 {
 496         ssize_t err;
 497
 498         if (mode == MCOPY_ATOMIC_CONTINUE) {
 499                 return mcontinue_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 500                                             wp_copy);
 501         }
 502
 503         /*
 504          * The normal page fault path for a shmem will invoke the
 505          * fault, fill the hole in the file and COW it right away. The
 506          * result generates plain anonymous memory. So when we are
 507          * asked to fill an hole in a MAP_PRIVATE shmem mapping, we'll
 508          * generate anonymous memory directly without actually filling
 509          * the hole. For the MAP_PRIVATE case the robustness check
 510          * only happens in the pagetable (to verify it's still none)
 511          * and not in the radix tree.
 512          */
 513         if (!(dst_vma->vm_flags & VM_SHARED)) {
 514                 if (mode == MCOPY_ATOMIC_NORMAL)
 515                         err = mcopy_atomic_pte(dst_mm, dst_pmd, dst_vma,
 516                                                dst_addr, src_addr, page,
 517                                                wp_copy);
 518                 else
 519                         err = mfill_zeropage_pte(dst_mm, dst_pmd,
 520                                                  dst_vma, dst_addr);
 521         } else {
 522                 err = shmem_mfill_atomic_pte(dst_mm, dst_pmd, dst_vma,
 523                                              dst_addr, src_addr,
 524                                              mode != MCOPY_ATOMIC_NORMAL,
 525                                              wp_copy, page);
 526         }
 527
 528         return err;
 529 }
 530
 531 static __always_inline ssize_t __mcopy_atomic(struct mm_struct *dst_mm,
 532                                               unsigned long dst_start,
 533                                               unsigned long src_start,
 534                                               unsigned long len,
 535                                               enum mcopy_atomic_mode mcopy_mode,
 536                                               atomic_t *mmap_changing,
 537                                               __u64 mode)
 538 {
 539         struct vm_area_struct *dst_vma;
 540         ssize_t err;
 541         pmd_t *dst_pmd;
 542         unsigned long src_addr, dst_addr;
 543         long copied;
 544         struct page *page;
 545         bool wp_copy;
 546
 547         /*
 548          * Sanitize the command parameters:
 549          */
 550         BUG_ON(dst_start & ~PAGE_MASK);
 551         BUG_ON(len & ~PAGE_MASK);
 552
 553         /* Does the address range wrap, or is the span zero-sized? */
 554         BUG_ON(src_start + len <= src_start);
 555         BUG_ON(dst_start + len <= dst_start);
 556
 557         src_addr = src_start;
 558         dst_addr = dst_start;
 559         copied = 0;
 560         page = NULL;
 561 retry:
 562         mmap_read_lock(dst_mm);
 563
 564         /*
 565          * If memory mappings are changing because of non-cooperative
 566          * operation (e.g. mremap) running in parallel, bail out and
 567          * request the user to retry later
 568          */
 569         err = -EAGAIN;
 570         if (mmap_changing && atomic_read(mmap_changing))
 571                 goto out_unlock;
 572
 573         /*
 574          * Make sure the vma is not shared, that the dst range is
 575          * both valid and fully within a single existing vma.
 576          */
 577         err = -ENOENT;
 578         dst_vma = find_dst_vma(dst_mm, dst_start, len);
 579         if (!dst_vma)
 580                 goto out_unlock;
 581
 582         err = -EINVAL;
 583         /*
 584          * shmem_zero_setup is invoked in mmap for MAP_ANONYMOUS|MAP_SHARED but
 585          * it will overwrite vm_ops, so vma_is_anonymous must return false.
 586          */
 587         if (WARN_ON_ONCE(vma_is_anonymous(dst_vma) &&
 588             dst_vma->vm_flags & VM_SHARED))
 589                 goto out_unlock;
 590
 591         /*
 592          * validate 'mode' now that we know the dst_vma: don't allow
 593          * a wrprotect copy if the userfaultfd didn't register as WP.
 594          */
 595         wp_copy = mode & UFFDIO_COPY_MODE_WP;
 596         if (wp_copy && !(dst_vma->vm_flags & VM_UFFD_WP))
 597                 goto out_unlock;
 598
 599         /*
 600          * If this is a HUGETLB vma, pass off to appropriate routine
 601          */
 602         if (is_vm_hugetlb_page(dst_vma))
 603                 return  __mcopy_atomic_hugetlb(dst_mm, dst_vma, dst_start,
 604                                                src_start, len, mcopy_mode,
 605                                                wp_copy);
 606
 607         if (!vma_is_anonymous(dst_vma) && !vma_is_shmem(dst_vma))
 608                 goto out_unlock;
 609         if (!vma_is_shmem(dst_vma) && mcopy_mode == MCOPY_ATOMIC_CONTINUE)
 610                 goto out_unlock;
 611
 612         /*
 613          * Ensure the dst_vma has a anon_vma or this page
 614          * would get a NULL anon_vma when moved in the
 615          * dst_vma.
 616          */
 617         err = -ENOMEM;
 618         if (!(dst_vma->vm_flags & VM_SHARED) &&
 619             unlikely(anon_vma_prepare(dst_vma)))
 620                 goto out_unlock;
 621
 622         while (src_addr < src_start + len) {
 623                 pmd_t dst_pmdval;
 624
 625                 BUG_ON(dst_addr >= dst_start + len);
 626
 627                 dst_pmd = mm_alloc_pmd(dst_mm, dst_addr);
 628                 if (unlikely(!dst_pmd)) {
 629                         err = -ENOMEM;
 630                         break;
 631                 }
 632
 633                 dst_pmdval = pmd_read_atomic(dst_pmd);
 634                 /*
 635                  * If the dst_pmd is mapped as THP don't
 636                  * override it and just be strict.
 637                  */
 638                 if (unlikely(pmd_trans_huge(dst_pmdval))) {
 639                         err = -EEXIST;
 640                         break;
 641                 }
 642                 if (unlikely(pmd_none(dst_pmdval)) &&
 643                     unlikely(__pte_alloc(dst_mm, dst_pmd))) {
 644                         err = -ENOMEM;
 645                         break;
 646                 }
 647                 /* If an huge pmd materialized from under us fail */
 648                 if (unlikely(pmd_trans_huge(*dst_pmd))) {
 649                         err = -EFAULT;
 650                         break;
 651                 }
 652
 653                 BUG_ON(pmd_none(*dst_pmd));
 654                 BUG_ON(pmd_trans_huge(*dst_pmd));
 655
 656                 err = mfill_atomic_pte(dst_mm, dst_pmd, dst_vma, dst_addr,
 657                                        src_addr, &page, mcopy_mode, wp_copy);
 658                 cond_resched();
 659
 660                 if (unlikely(err == -ENOENT)) {
 661                         void *page_kaddr;
 662
 663                         mmap_read_unlock(dst_mm);
 664                         BUG_ON(!page);
 665
 666                         page_kaddr = kmap_local_page(page);
 667                         err = copy_from_user(page_kaddr,
 668                                              (const void __user *) src_addr,
 669                                              PAGE_SIZE);
 670                         kunmap_local(page_kaddr);
 671                         if (unlikely(err)) {
 672                                 err = -EFAULT;
 673                                 goto out;
 674                         }
 675                         flush_dcache_page(page);
 676                         goto retry;
 677                 } else
 678                         BUG_ON(page);
 679
 680                 if (!err) {
 681                         dst_addr += PAGE_SIZE;
 682                         src_addr += PAGE_SIZE;
 683                         copied += PAGE_SIZE;
 684
 685                         if (fatal_signal_pending(current))
 686                                 err = -EINTR;
 687                 }
 688                 if (err)
 689                         break;
 690         }
 691
 692 out_unlock:
 693         mmap_read_unlock(dst_mm);
 694 out:
 695         if (page)
 696                 put_page(page);
 697         BUG_ON(copied < 0);
 698         BUG_ON(err > 0);
 699         BUG_ON(!copied && !err);
 700         return copied ? copied : err;
 701 }
 702
 703 ssize_t mcopy_atomic(struct mm_struct *dst_mm, unsigned long dst_start,
 704                      unsigned long src_start, unsigned long len,
 705                      atomic_t *mmap_changing, __u64 mode)
 706 {
 707         return __mcopy_atomic(dst_mm, dst_start, src_start, len,
 708                               MCOPY_ATOMIC_NORMAL, mmap_changing, mode);
 709 }
 710
 711 ssize_t mfill_zeropage(struct mm_struct *dst_mm, unsigned long start,
 712                        unsigned long len, atomic_t *mmap_changing)
 713 {
 714         return __mcopy_atomic(dst_mm, start, 0, len, MCOPY_ATOMIC_ZEROPAGE,
 715                               mmap_changing, 0);
 716 }
 717
 718 ssize_t mcopy_continue(struct mm_struct *dst_mm, unsigned long start,
 719                        unsigned long len, atomic_t *mmap_changing)
 720 {
 721         return __mcopy_atomic(dst_mm, start, 0, len, MCOPY_ATOMIC_CONTINUE,
 722                               mmap_changing, 0);
 723 }
 724
 725 void uffd_wp_range(struct mm_struct *dst_mm, struct vm_area_struct *dst_vma,
 726                    unsigned long start, unsigned long len, bool enable_wp)
 727 {
 728         struct mmu_gather tlb;
 729         pgprot_t newprot;
 730
 731         if (enable_wp)
 732                 newprot = vm_get_page_prot(dst_vma->vm_flags & ~(VM_WRITE));
 733         else
 734                 newprot = vm_get_page_prot(dst_vma->vm_flags);
 735
 736         tlb_gather_mmu(&tlb, dst_mm);
 737         change_protection(&tlb, dst_vma, start, start + len, newprot,
 738                           enable_wp ? MM_CP_UFFD_WP : MM_CP_UFFD_WP_RESOLVE);
 739         tlb_finish_mmu(&tlb);
 740 }
 741
 742 int mwriteprotect_range(struct mm_struct *dst_mm, unsigned long start,
 743                         unsigned long len, bool enable_wp,
 744                         atomic_t *mmap_changing)
 745 {
 746         struct vm_area_struct *dst_vma;
 747         unsigned long page_mask;
 748         int err;
 749
 750         /*
 751          * Sanitize the command parameters:
 752          */
 753         BUG_ON(start & ~PAGE_MASK);
 754         BUG_ON(len & ~PAGE_MASK);
 755
 756         /* Does the address range wrap, or is the span zero-sized? */
 757         BUG_ON(start + len <= start);
 758
 759         mmap_read_lock(dst_mm);
 760
 761         /*
 762          * If memory mappings are changing because of non-cooperative
 763          * operation (e.g. mremap) running in parallel, bail out and
 764          * request the user to retry later
 765          */
 766         err = -EAGAIN;
 767         if (mmap_changing && atomic_read(mmap_changing))
 768                 goto out_unlock;
 769
 770         err = -ENOENT;
 771         dst_vma = find_dst_vma(dst_mm, start, len);
 772
 773         if (!dst_vma)
 774                 goto out_unlock;
 775         if (!userfaultfd_wp(dst_vma))
 776                 goto out_unlock;
 777         if (!vma_can_userfault(dst_vma, dst_vma->vm_flags))
 778                 goto out_unlock;
 779
 780         if (is_vm_hugetlb_page(dst_vma)) {
 781                 err = -EINVAL;
 782                 page_mask = vma_kernel_pagesize(dst_vma) - 1;
 783                 if ((start & page_mask) || (len & page_mask))
 784                         goto out_unlock;
 785         }
 786
 787         uffd_wp_range(dst_mm, dst_vma, start, len, enable_wp);
 788
 789         err = 0;
 790 out_unlock:
 791         mmap_read_unlock(dst_mm);
 792         return err;
 793 }